HEAgent: Autonomous Large Language Model-Driven Framework for High-Entropy Alloy Electrocatalyst Discovery

High-entropy alloys (HEAs) are promising electrocatalysts for the oxygen reduction reaction (ORR), as they offer diverse local adsorption environments for tuning catalytic activity. However, their vast composition spaces make exhaustive density functional theory (DFT) or experimental screening impractical, motivating the development of efficient optimization strategies. We present HEAgent, a large language model (LLM)-driven multi-agent workflow that proposes, simulates, and iteratively refines HEA compositions for ORR. In a quinary composition space defined over 15 candidate elements, the best-performing HEAgent variant attained a mean best Pt(111)-normalized activity of 4.64, exceeding the Bayesian optimization baseline by 21% and the Random baseline by 62% under equal proposal budgets. Trajectory analysis of this space showed that HEAgent initially proposed higher-activity Pd/Pt-rich compositions than the randomly initialized baselines, followed by subsequent feedback-guided refinement within these families. In a separate search in the restricted IrPdPtRhRu space, the highest-activity composition observed was recovered in seven of ten HEAgent repeats but in none of ten baseline repeats, indicating more consistent recovery by HEAgent. These results show that LLM-guided proposal generation and iterative feedback can improve sampling of HEA compositions under limited evaluation budgets.